Professional Documents
Culture Documents
WORKING PAPER
ALFRED P. SLOAN SCHOOL OF MANAGEMENT
by
MASSACHUSETTS
INSTITUTE OF TECHNOLOGY
50 MEMORIAL DRIVE
CAMBRIDGE, MASSACHUSETTS 02139
THE EMERGENCE OF A NEW TECHNOLOGY:
THE CASE OF NEURAL NETWORKS
by
Michael Rappa is assistant professor of management, in the Management of Technology Group at the MIT
Sloan School of Management. Koenraad Debackere is a visiting fellow at Sloan and a doctoral candidate in
management at Ghent University in Belgium. The authors are particularly grateful to Susan Bj0mer of the MIT
Library for her expert assistance in electronic database searching, and to Professor Thomas Allen for his advice
and encouragement This research was sponsored, in part, by the Swedish Board of Technical Development
Mf T . (00«OICC I
THE EMERGENCE OF A NEW TECHNOLOGY:
THE CASE OF NEURAL NETWORKS
by
June 1989
Abstract
Introduction
This paper explores the dynamic characteristics of R&D communities. The notion "R&D
community" is used to signify a group of individuals, composed of scientists or engineers,
who are committed to solving a set of interrelated scientific and technological problems, who
may be organizationally and geographically dispersed, and who communicate in some way
with each other, i The ultimate goal for some members of the community may be to create new
knowledge, while for others it may be to apply existing knowledge in the creation of new
products or processes. Furthermore, the community can include individuals employed in any
corporations, and government research institutes wherever they may be located throughout the
worid.
This concept of an R&D community extends beyond traditional disciplinary and industry
boundaries. Although it is common to speak of entities such as the "physics community," this
is not what is intended by the term R&D community; nor is it meant to be a finer grain subset
of researchers within a discipline, such as those involved in solid state physics. Rather, the
community is normally interdisciplinary in nature and can include researchers from a wide
variety of academic specialties. Moreover, within the private sector, it includes firms
irrespective of their standard industrial classification. The only requirement is their focus on
some part of the relevant set of problems. Thus, the community is defined here by the nature
of the problem set and not necessarily by the end product, as is normally the case with an
industry definition.
'The notion of a research community has been developed extensively by scholars who have focused strictly on
academic scientists and the so-called "invisible colleges" which arise in various problem areas (ZLman, 1984). The
R&D community proposed in this paper seeks to extend the concept of an invisible college, to offer a more
encompassing view, and in doing so, explore its relevance to technological development.
The dynamics underlying R&D communities are explored by studying the example of the
neural networks community. This paper seeks to broaden the scope of current thinking on
technological development which tends to emphasize the instrumental role of the firm in
bringing new technology to the marketplace. Much academic research on the subject of
technological development has been directed at understanding the functioning of firms (and
doubt that previous work along this line of inquiry has yielded important insights about the
management of technology development, but in focusing on the firm in isolation of the broader
environment in which technological development occurs, it may have lost sight of some other
firms—or for that matter, a collection of firms in a given industry— but rather an activity which
cuts across many types of public and private organizations. Furthermore, it is proposed that a
new technological development may come about through the concerted efforts of a community
of researchers which forms over time and that spans these diverse organizations. If this is
indeed the case, then the progress of such communities may be better understood through a
careful examination of how they function. This functioning is now further illustrated in the
processes, which has certain features that set it apart from the traditional and most widely used
^This study of neural networks is part of a larger study of the role of R&D conununities in the emergence of new
technologies.
type of information processing system, the digital computer based on the von Neumann
programmed in the usual sense, but rather it is trained with data. This is an attractive feature
because some information processing tasks are cumbersome from a programmer's perspective,
and indeed devising the proper algorithm to perform a particularly complex task may be
virtually impossible. It also implies that a neural network benefits from experience: as it
processes more and more information while performing a task, it becomes increasingly more
The second notable feature of a neural network is its degree of parallelism in processing a
task. Unlike a traditional computer, with a single or small number of sophisticated central
processing units, a neural network has a very large number of simple processing elements that
parallelism is in its processing speed and fault tolerance (i.e., its ability to perform a task even
though some processing elements fail.). In sum, although they may vary in form and
simple (usually adaptive) elements and their hierarchical organizations which are intended to
interact with the objects of the real world in the same way as biological nervous systems do"
(Kohonen, 1988).
A clear distinction exists between neural computing and digital computing. Training a
computer with data, or having it learn, are concepts quite foreign to the normal operation of the
modem digital computer. The basic computer consists of a storage unit into which data are fed
and a central processing unit with circuitry that can perform certain logic functions. It operates
^The neural network field is also referred to as "connectionism," "adaptive systems," or "neurocomputing." For a
comprehensive overvievk- of neural network technology, see the DARPA Neural Network Study (1988), or for a brief
review see Hecht-Nielsen (1988).
in accordance with a stored program (or set of algorithms) that retrieves, manipulates, and
stores bits of data (in binary form) in the course of its computational tasks. As Kohonen
(1988) points out, biological neural systems do not apply principles of digital or logic circuits
and therefore, the collective processes which are important in neural computing simply cannot
machine instructions or control codes occur in neural computing. The circuits in neural
networks do not implement recursive computation and are thus not algorithmic. Nor is the
neural system's memory storage localized in a particular unit; instead, the memory function is
distributed among the numerous processing elements. For this reason, Hecht-Nielsen (1988)
The unique characteristics of neural networks may make them well-suited for certain
classes of processing tasks that are otherwise difficult to perform using traditional
computational techniques, and vice versa. For example, much like the human brain, neural
networks are expected to perform best in applications requiring the processing of sensory
information and controlling interactions with the environment, such as speech recognition,
machine vision, robotic control, signal processing, and pattern recognition. But conversely,
neural networks perform poorly in applications where rule-based problem solving can be
Neural networks have a long history of development. Early research has its roots in
theoretical explanations of the brain and thinking processes proposed during the 1940s.
During the initial decades, attention focused on the fundamentals of neural computing,
including path breaking work on the formulation and elaboration of basic models to explain
such phenomena as adaptive stimulus-response relations in random networks. The 1960s
neural network research, but for reasons which are discussed at greater length below, the work
controversy that ensued resulted in a decline in government support of neural network research
Nonetheless, work on neural networks continued during the 1970s, with a small number
of dedicated researchers making significant progress on the theoretical foundations. The field
developments had been made in both theory and application. Three general factors appear of
paramount importance in explaining the recent growth of the field: (1) the evolution of the
single-layer perceptron into a multi-layer system; (2) the rapid development of other
simulate, and diagnose neural nets of greater sophistication; and (3) significant progress in the
advances in semiconductor integrated circuit technology, Sivilotti, Emerling, and Mead of the
California Institute of Technology designed a VLSI collective decision circuit in 1985. This
development and the subsequent designs of silicon models for neural computation set the stage
for the design of neural network hardware. Even today, most neural nets are simulated on
digital computers. However, the design of silicon models for neural computation brings the
There are a number of neural networks that have been in the process of development during
the last few years and several are now available in the marketplace. Among those neural
networks marketed commercially, some are software implementations that run on digital
computers, others are hardware implementations (neurocomputers), and still others are a
mixture of software and hardware. Neural networks are being explored for use in applications
as diverse as Kanji character recognition, risk analysis in bank lending, and predicting process
yields on a manufacturing line in real time. Although first generation products have attracted
much attention, it is the promise of the next generation which is capturing the imagination of
potential seems enormous, though many problems still have to be solved before the technology
reaches its full potential. The following sections of this paper examine the historical evolution
of neural network community, seeking to shed light on the structural and behavioral
bootlegging, and the grapevine. We also examine the controversy that arises between different
communities, and the institutionalizing mechanisms used to preserve a community's
momentum as it emerges. Given the experience of the neural network community, we will
then discuss the general implications for a model for understanding the emergence of radically
new technologies.
It is proposed that the emergence of a new technology is to a large extent the work of a
dedicated R&D community. The community is formed and nurtured not by administrative
decree, but by the autonomous actions of individual researchers who become intrigued by an
idea and are committed to solving the problems necessary to make that idea work. In a sense,
it is a self-organizing process. Moreover, it is a subtle form of collective action in that it
naturally arises in pursuit of the rewards that will accrue to those who arrive at new knowledge
and its application first. But even though researchers are competing with one another, they
may nonetheless see themselves as members of a larger community fi-om which they can draw
on, and contribute to, in a variety of ways. As will be illustrated below, it is this combination
of cooperation and competition that lies at the heart of R&D communities. This proposition is
Methodology
The data presented in this paper were obtained by studying the extensive body of scientific
and technical literature generated by the neural network researchers themselves in the course of
their work. Since communication among members is a defined characteristic of the R&D
community, studying the documented, or formal, communication (in the form of papers and
written presentations) among researchers is a convenient means for gaining insight into the
information about the technology itself, and subsequently about the behavior and the structure
of the neural network community. In order to further systematize the study of the neural
networks community, an electronic relational database was analyzed containing 2740 abstracts
of neural network articles drawn from journals and conference proceedings since 1969. This
analysis does not focus, however, on the well-established bibliometric methods for analyzing
citation frequency and co-citation clusters. Instead, the abstracts were used in a different
manner: namely, for the detailed information contained within each document.
Indeed, each abstract provides a wealth of reliable information about research activities
within the community. The first advantage is its longitudinal character. By studying an R&D
community over a period of years, it becomes possible to show the dynamic patterns which
emerge. The second advantage lays in the information obtained from the abstracts. Each
abstract tells us who the researchers are and how their numbers vary from year-to-year. The
abstracts also show what topics the researchers are working on and which organizations
employ them. They also reveal the ties which develop over time among the different
researchers and organizations by looking at the co-authorship of the papers and by showing the
a two-decade time span, one is able to visualize the structural changes in the R&D community,
and also, to draw inferences about the behavior of the researchers dedicated to the particular
technology.
Be that as it may, the literature does have certain non-trivial limitations. First, formal
likely occurs among researchers at conferences and via telephone and computer networks.
Partly as a consequence, individuals participating in the community may not become visible in
the documented literature. Furthermore, the communication that does get pubhshed is slightly
delayed in the process and not necessarily reflective of all the important details of the research.
There may also be activity in developing the technology that is purposely hidden from public
Nevertheless, given these limitations, the literature can prove to be remarkably useful in
to ascribe more or less importance to a particular publication, and by analyzing the data
historically. Thus, for example, we are not seeking to understand absolute magnitudes so
much as the dynamic trends over time, and we are not seeking to uncover technical secrets so
much as obtaining a basic understanding of the people involved, the nature of their work, who
they worked with, where they worked, and when they worked. Taken together, the data
culled from the literature can provide a comprehensive picture of change over time within the
community.
An analysis of the neural networks database was conducted to determine the number of
individual researchers worldwide, who were active each year in the research and development
author from year-to-year. In this way, membership in the community is not sensitive to
number of publications by an author in a given year. This analysis yields a growth profile of
the neural networks community between 1969 and 1988, which is shown in figure 1. The data
indicate that until 1985, the community was rather small, with no more than 200 researchers
active annually. However, after 1985 the community expanded very rapidly, from about 200
researchers to over 1000, which translates into a compound growth rate of over 60-percent.
During this phase the community blossomed into a disciplinary collage composed of
that is, the rush to join in on a trend, which has less to do with the underlying fundamentals
than with the propensity for people to act on the basis of the actual or perceived actions of
V
M
u
68 70 72 74 76 78 80 82 84 86 88
Year
new technology, a race for the rewards that accrue to those who are first to stake claims to new
knowledge in the form of patents or papers. The result is the same regardless of the
motivations of the individual researchers involved, whether it be fame, fortune or fad: a chain
other. It is interesting to note that in the case of neural networks, we found the bandwagon-
effect to occur similarly in both public and private sectors and in different countries.
An influential member of the neural networks community. Carver Mead, observes: "In a
field like neural networks, one is usually too optimistic in the short run, but one is never
optimistic enough in the long run" (DARPA, 1988). The bandwagon-effect may be fueled by
an excessive degree of optimism among researchers, and may ultimately lead to the
11
proliferation of hyperbole in the form of unfounded and unrealistic expectations about the
technology's potential. This is particularly true in conjunction with the mass media, where the
neural networks community has received its fair share of "hype." The DARPA report
addresses this issue directly, stating that the neural networks field lends itself readily to bold
claims and sensational media headlines, such as proclamations of "thinking computers," and
concludes that it is unlikely to dissipate: "Hyperbole's natural habitat is, after all, the territory
between achievement and promise: the greater the potential of a new and mostly unexplored
avenue of research, the greater the human tendency to imagine it opening pathways to both
The proliferation of hype runs against the aspired norms of researchers, who as a group
generally despise self-promoting hucksters who, in search for financial support, dare to
oversell the potential of their research. For example, the founder of one neural networks
company claims, "There is a whole lot of hype and nonsense in the field, which takes away
from the credibility of the people who have been doing legitimate work. There has been
important progress, but some people are grossly over portraying what the machines can do
But, while it may well be contemptible to some, hype is a relative concept, and to a certain
extent it may have the positive effect of attracting peoples' attention, including managers,
venture capitalists, govemment officials-all of whom might have funds to allocate. Perhaps
most important, is the need to attract other researchers to the field. Although neural network
researchers may shrink when they see overly optimistic assessments of their field in print, they
also understand the need for others (including students) to join in and become seriously
committed to solving the difficult problems which confront them. The challenges may simply
be too immense to be realistically accomplished in the near future unless the community attains
12
a certain level of panicipation, which researchers refer to as critical mass. For example, a
leading academic researcher in neural networks and director of the DARPA study states, "I
believe all the hype. What worries me is that many people don't realize how hard the problems
are. Even if something works in the laboratory, it still takes ten years to get it out in the
marketplace (Kinoshita and Palevsky, 1987)." The trick lies in generating a healthy optimism
that creates enthusiasm rather than alienation among potential entrants to the community.
knowledge, in combination with a tendency toward optimism, yielded a very dramatic rise in
the level of participation in the neural network community indicative of the bandwagon-effect.
If the field is fertile and matches its promise, perhaps it will continue to enjoy a healthy
expansion in participation. But what will happen if researchers discover, in the course of their
experiments, that the problems they face are in fact much too difficult to resolve in the
foreseeable future? Or that other technologies appear to be more promising? If researchers are
so fast to enter a field, is it also likely that they might be quick to leave because their overly
reveals a pattern that is suggestive of the cycles of enthusiasm and despair that might befall a
research community. Figure 2 shows the growth of the neural network community over the
sixteen year period from 1970 to 1985, and makes a distinction between the core group of
researchers (those who are usually active from year-to-year) and the rest of the community.
Specifically, the core group is defined to include researchers who were actively participating in
the community in at least three years of a given five-year period. Notice that the core group is
relatively small (on average, between ten- to twenty-percent of the total community) and fairly
13
stable in size over the entire period, as should be expected. They are the researchers who keep
the faith regardless of the obstacles that lie ahead, the true believers in the technology even
when others disparage their work, the hopelessly committed. Success in their effort to push
the field forward will preserve their place as founding fathers and gamer the respect of their
colleagues. However, sadly, if progress is slow or elusive, they will suffer a reputation as
Figure 2 also illustrates the cyclical pattern of participation among researchers whose
commitment is less enduring than the core group. It is apparent from the data that the majority
of researchers active in any given year are involved only briefly. Why this pattern arises is
open to speculation. Perhaps it derives from the tendency for researchers to dapple in different
areas which are tangential to their primary area of interest. It also might arise from the tradeoff
researchers make in judging the likely potential of succeeding with one particular avenue of
250
-
200
Total Community
(R^2 = 0.88)
-
13 150
68 70 72 74 76 7J 82 84 86
Year
research versus another. Or it might to some extent occur as master's and doctoral students
graduate and pursue other interests. In any case, the data suggest a degree of fluidity of
participation in neural network community that stems from the movement of researchers
between the variety of problem areas, given their expertise, they might pursue.
This fluidity of researchers into and out of the community is shown in greater detail in
Figure 3. The database was partitioned into five, three-year periods in order to examine the
new entrants (those not active in the previous period), sustainers (those also active in the
previous period) and exits (those who were active in the last period by not in the present one).
The degree of turnover from period to period is quite remarkable, with many researchers
800
600
400-
«i 200-
-200 -
-400
15
An analysis of the extent to which neural network researchers who having left in one
period return to the community in a later period was also conducted. It indicates that between
10- and 20-percent of entrants in a period (three through five) were previously involved in the
community.
The foregoing analysis suggests that R&D communities, in their emergent stage, have a life
of their own largely free of the formality that comes from organized programs of research,
informality and dynamism that seemingly derives from the independent decisions of
researchers in search of promising new technological frontiers. How else could a community
grow so rapidly or researchers move so fluidly, if not otherwise uninhibited by the red-tape
usually associated with organizational decision-making in the public or private sector? But if
researchers do act independently, how do they acquire the resources to sustain their effort?
The astounding growth of the neural networks community cannot be ascribed to a dramatic
increase in public funding for neural network research. Although funding will certainly
increase in the coming years (this appears from the intentions of the DARPA study), it was not
likely the trigger for the dramatic increase in the number of researchers which occurred after
1985. Instead, the neural network community sustained itself and fostered its emergence not
so much from formal public and private sector funding, as from dedicated researchers working
more or less in the "shadows" of the laboratory. The terminology common among researchers
is bootlegging, a word used to describe the practice of scrapping together resources to work,
more or less surreptitiously, on interesting problems not sanctioned under existing budgets. It
might take the form of working during lunch, evenings, or weekends, or taking advantage of
the slack resources that typically exist in laboratories, or even using one's personal resources.
16
The practice of bootlegging is probably as old as the research laboratory itself, and seems
to be a central element of every good story retold by researchers about their heroic effort to
succeed in developing a radically new technology in spite of the powers that be. Bootlegging
nature of research in new technologies. Important research is very difficult to judge in its early
stages and thus, it is more likely to be rejected in favor of less risky, incremental work when
formally evaluated. Researchers dedicated to a new and unorthodox technology are confronted
with a difficult dilemma: they need more proof that their work will yield results before
receiving resources, but without resources they are unable to do precisely that. Bootlegging
enables fledgling research to go forward without the full knowledge and scrutiny of managers
and other researchers, to a point at which the promise of the idea is clear.
sponsorship of research presented at the First Annual International Neural Network Society
meeting held in Boston in 1988. Of the 527 papers presented, 324 (61%) mentioned no
sponsor at all. Only 19 papers (4%) reported research which was at least partly funded by the
National Science Foundation. We found 43 papers (8%) that were at least partly funded by
one of the US Defense Department agencies. Corporate involvement occurred in 108 instances
(20%). One paper even described itself as "privately funded research," without any ties with
the author's "current affiliations." One participant to the conference observes that in terms of
travel money alone (perhaps a few million dollars), very little could be accounted for by
funding agencies. Thus, although public research funds may start pouring into the neural
network research laboratories, this kind of support clearly did not provide the trigger for the
The relevant lesson from the practice of bootlegging is that researchers are driven by
interesting problems, and not necessarily by what is looked upon favorably by funding
agencies. Funding does not create a field, but rather allows it to flourish. If researchers make
progress in a new area, the funding will hopefully follow suit. Whether it does indeed follow
suit depends upon the researchers' fight for legitimacy among their peers, who may hold
The fight for legitimacy has been a particularly intense struggle in the case of neural
to work on neural network technology throughout the sevennes— the so-called "wilderness
years"--even though little funding was forthcoming and the technical feasibility of their
concepts was heavily attacked by other researchers. At the heart of the controversy is the 1969
MIT researchers, Marvin Minsky and Seymour Papert, who seriously questioned the potential
To Minsky and Papert, neural networks were (and still are) technically uninteresting,
researchers, their treatise was seen as a blatantly underhanded attack to discredit neural
network research in order to sway government funding away from the field. A leading neural
network researcher recalls, "My impression was that Minsky and Papert defined the perceptron
narrowly enough that it couldn't do anything interesting. ...It looked like an attempt to show
that the perceptron was no good. It wasn't fair (Kinoshita and Palevsky, 1987)." In defense
For his part, Minsky also admits that the attack on neural networks may have been too ardent,
but defends his book claiming, "We were faced with a shortage of researchers in AI, and we
were watching people throwing away their lives on perceptrons.... You would have to be crazy
to be doing perceptrons when you could get so much more dramatic results with AI (Kinoshita
andPalevsky, 1987)."
The bitter controversy lingers on and is perhaps as intense today as in 1969. (In an excerpt
provided in Appendix A, Papert muses over the controversy in a recent and somewhat
humorous allegory.) Although those within the neural network community contend that the
limits posed in Minsky and Papert's book have by now been overcome, both authors remain
skeptical about the future of neural nets: "Indeed, Minsky and I. ..suggest that the entire
structure of recent connectionist theories might be built on quicksand: it is all based on toy-
sized problems with no theoretical analysis to show that performance will be maintained when
the models are scaled up to realistic size (Papert, 1988)." Indeed, Papert goes so far as to
myth" that fits with the current intellectual fad to move away from rationalism toward more
In response to these attacks, neural network proponents have launched their own
19
vociferous criticisms of their detractors, suggesting that interest in neural networks stems, in
part, from the disappointing results of Al research— this, despite the massive amount of funding
the field has received over the years. Thus, in their view, it is the failure of AI that is breathing
life into neural networks, and such diatribes as those of Minsky and Papert simply mask the
Despite the whirling controversy and despite the lack of funding for neural network
believe in the potential of neural networks. Their diligence during lean years eventually paid
off in the early eighties, as their ideas gained favor among a rapidly growing number of
scientists and engineers employed across a diversity of disciplines and industries. The
organizational and geographical spread of the neural networks community is examined in the
following section.
As described in the introductory section, R&D communities can extend beyond industrial
boundaries, such that technological development is not wholely the domain of firms. R&D
communities can consist of researchers employed by a variety of organizations, cutting across
20
university, industry, and government sectors. The neural networks case exemplifies this fact
quite clearly.
For our purposes, the organizations involved in neural network research were classified as
being either within the private or public sector. This second group is a mixture of university
researchers between the two sectors is shown in figure 4. Since 1969, a cumulative total of
about 3800 man-years of research has been devoted to the development of neural networks.
The fast majority of that effort took place in the public sector. Although industrial researchers
had grown in numbers during the bandwagon-phase (reaching 18-percent in the last few
years), prior to 1985 only between two- to seven-percent of all researchers where in the private
sector.
The neural network community is an excellent example of how the efforts of academic
researchers can contribute to the practical development of a new technology. Although neural
networks have their roots in an academic setting, they have emerged in the eighties as a
commercial activity. Many academic members of the community attempt to demonstrate the
commercial viability of their research, and some even participate in private organizations in a
variety of capacities. Like the more well known examples of semiconductor technology or
genetic engineering, neural networks suggest that science and technology can and does become
The neural network community is also quite international in scope, with more than a dozen
countries participating. Indeed, researchers from relatively smaller industrial countries, such
through the seventies and now figure prominently in the field. Figure 5 shows
21
1000
Private Sector
800-
u D Public Sector
600-
u
400-
I 200
I r t I 'l
'
69 71 73 75 77 79 81 83 85 87
Year
the international distribution of researchers among the six major industrial countries (the United
Kingdom, West Germany, France, Italy, Japan, and the United States^) involved in the
development of neural networks. In the post- 1985 bandwagon phase, approximately seventy-
to eighty-percent of the community resided in these six countries. It is interesting to note that
the distribution among these countries varies over the entire period from 1969 to 1988, but that
since 1985, the U.S. share has dominated the community, with about sixty-percent of the total.
Thus, the rapid growth of the neural network community has been centered most strongly in
the U.S.
As the neural network community grew in the period from 1982 onward, it became more
widely diffused across public and private sector organizations. One way of viewing the extent
''According to the NSF, these six countries account for about 85-percent of the world's total research and
development expyenditures (NSF, 1989).
22
23
100.0 « • —
o Private Sector
•a
75.0 -
6
m
a.
o 50.0 -
f-
3 =
Z .5
-
25.0
o u
c o
S E
0.0 J I 1 1 i_
81 83 87 89
Year
Although the public sector community had consistently been more widely diffused than the
private sector, it nevertheless also experienced a similar increase in the number of participating
university and government laboratories, from about 25 in 1982 to nearly 250 in 1988. Thus,
in both sectors, we found a rapid diffusion of the community across different organizations.
Simultaneous with the growth and diffusion of the community, we see a tremendous
increase in the number of firms that for the most part specialize in the development and
application of neural networks. These finns are typically small in size, and in many cases they
are ventures formed specifically for exploiting neural network technology. Investigation of the
database led to the identification of 36 such firms. It appears that these firms form a bridge in
the technology's transition from the public to the private sector. It is commonplace to find
academic researchers deeply involved with the new ventures, taking part in their incorporation.
24
The rapid growth and diffusion of the community likely has a beneficial effect on the
array of organizations, seeking to develop and apply the technology in a increasing variety of
ways, should almost certainly contribute further to the advances already made. As the
community expands and spreads, it develops a powerful momentum that derives from the force
of its numbers and the ingenuity of researchers independently working in laboratories in every
However, the advantages of this expansion come with a cost: the diminishing ability of
researchers to easily communicate with one another. This is a heavy cost, indeed, since it is
the communication of information and knowledge among researchers that holds a community
together. As a community grows in size, it quickly becomes a virtual impossibility for any
given researcher to frequently communicate information to all others without some degree of
greater length below, is through the published literature. However, this is a very rudimentary
communication mechanism with severe limitations given the constant flow of new, complex
grapevine.
The grapevine is an informal, but remarkably efficient network of researchers who facilitate
the flow of information among different laboratories. Chances are the researchers are well
acquainted with each other, perhaps having previously worked or studied together, or having
become ftiends at a conference while commiserating over the years spent toiling over similar
25
problems. Moreover, the core researchers discussed above are likely central nodes in the
grapevine. Indeed, over the two decades which have passed since 1969, a highly
interconnected network among the different organizations involved in neural network research
has developed. The network, illustrated in Figure 7, was analyzed by examining the database
for instances where researchers in two different organizations were co-authors, or where a
particular researcher (permanently or temporarily) changed his laboratory affiliation over the
twenty-year period. The obvious assumption here is that researchers who collaborate on a
paper are also likely to talk with one another even though they are not (or no longer) working
in the same laboratory. In any case, this is possibly one of the most stringent criteria one can
use to define a link between two organizations, such that the true grapevine is likely to be far
more elaborate.
appears that is likely representative of the core structure of the grapevine. The neural network
grapevine, which in total connects 1 15 organizations, provides further supporting evidence for
our basic proposition of the existence of R&D communities. Over the twenty year period, a
total of 568 different organizations have been identified in the database. Thus, 20-percent of
them are connected in the core grapevine. It is interesting to note that industrial research
laboratories frequently appear in the central nodes. Moreover, while U.S. and European
organizations appear fairly well integrated, Japanese organizations are less well connected to
the core and some are part of a second, smaller network which is disconnected from the core.
Further analysis indicates that the complexity of the neural network grapevine has evolved
over time and become particularly pronounced in the bandwagon phase of the community's
maturation. In the pre- 1985 era, very few organizations were linked together, and those that
were constituted mostly of simple, bilateral linkages. It was not until after 1985 that the first
26
NON-INIXJSTRY
semblance of a more far-reaching grapevine became evident, and in the last year alone it has
doubled in density.
dynamics that may hold special significance in understanding the emergence of a new
technology. We have seen a community, which after years of wavering enthusiasm and
despair, eventually take root, grow rapidly, and spread across hundreds of organizations in the
public and private sector. While at the same time the community maintained its contiguity,
27
preserving the flow of information among its members via a rather extensive grapevine. But
this remarkable surge of activity is not alone sufficient to ensure the success of the neural
network community in commercializing its technology. The R&D community must continue to
sustain a high level effort, or, in the vernacular of researchers, preserve its momentum. In the
following section we describe some of the basic, and fairly obvious, institutions that evolve
Institutionalizing meclianisms
main purpose of these institutions is to ensure the community's healthy development in the
post-bandwagon phase, by structuring the entry of new researchers into the field, easing the
exchange of knowledge and information among members, and working to secure the funds
necessary to sustain the effort long enough until the technology becomes commercially viable.
First of all, formal professional societies get established. Thus, the neural networks
community created its "International Neural Network Society" (INNS), which now lists 3,500
members fi-om 38 countries (with students comprising over one-fifth of the membership). The
primary objective of such societies is to facilitate the transfer of information and knowledge
and lecture series, and sponsoring the publication of a journal. In the words of its president,
the Society works as an "interdisciplinary catalyst," with the goal of providing "information,
engineering (Widrow, no date). Toward this end, the INNS created its journal Neural
Networks in 1988 and organizes an annual conference. (Another publication, the Journal of
Neural Network Computing: Technology, Design and Applications , has been established
28
independent of the INNS.) The force behind these institutions tends to be the core group of
researchers earlier identified, who serve as officers of the society and who sit on the editorial
Beside allowing for the presentation of recent work, conferences provide an opportunity
for the interpersonal exchanges that aid significantly in the solidification of collegial
relationships— key to the functioning of the grapevine when researchers return to the
laboratory. Between 1988 and early 1989, we were able to identify thirty-one conferences
which were partly or totally devoted to neural networks technology. Perhaps the premier
conference was the First Annual INNS Meeting in Boston in 1988, with over 600 papers
Another mechanism serving to speed the communication process, are simple newsletters
that are written to quickly disseminate the latest news on progress in the field. They can be in
the form of a brief booklet (like the Neural Network Review), or even a "gossip sheet" (such
with a penchant for scuttlebutt. One variant on this is the firm-sponsored newsletter, such as
Neuralware's Connections, which serves more as a means for publicizing products and recent
applications.
Perhaps the speediest and most advanced mechanism for exchanging information is the
communication network system that links laboratory computers across the world. In
particular, the invention of the "interest group" or "bulletin board" format on these networks
provides an astonishingly fast and efficient means for mass communication among members of
the community. Although it is a electronic medium, like the telephone and the facsimile
machine, the computer network interest group is more like the traditional printed publication in
29
its ability to reach large numbers of researchers simultaneously, only much more quickly. A
neural networks interest group has already been in operation for at least a year, and the INNS
Another important institutionalizing mechanism, used primarily for the codification and
transfer of the community's knowledge base, is the textbook. A number of neural network
textbooks have appeared (among leading publishers is the MIT Press, with already nine
textbooks in print). Textbooks also play a central role in the education of new recruits into the
community, especially when woven into a coherent curriculum. This points to yet another
mechanism, the formal graduate research program, which provides a direct means for
channeling young researchers into the community and ensures that knowledge of the field will
universities have started offering graduate programs focused on neural networks. Already
seven in number, these universities are highly central in the community grapevine. Along with
the formation of graduate programs, specialized research centers and laboratories are also
established. In this way, the community gains recognition as a separate entity within
university settings, just as it does within the corporate world (through the creation of
Lastly, there are structured efforts to ensure that funds will flow into the field in sufficient
quantity to support researchers who are committed to developing the technology. The DARPA
study, which was conducted with the help of more than fifty neural network researchers, is
just one example of the community's effort to gain financial sponsorship. There have also
been specific conferences and workshops dedicated to bringing together researchers with
potential sponsors, in order to educate funding sources about the potential of neural networks
There is very little novelty to the institutionalizing mechanisms we have observed in the
neural network community. Nonetheless, what is interesting is that these institutions emerged
from the grassroots of the community, with remarkable speed and strength, and without central
independently, but moving with a synchronism that derives from a shared goal. Taken
together, these institutions are quite effective in preserving the community's momentum.
Discussion
networks case suggests some basic elements that may contribute to a general theory of how the
emergence of a new technology occurs. In this section, we offer a number of propositions that
might fit within such a theory, given our observations of the structural and behavioral
characteristics of the neural networks R&D community. It is readily admitted that this
discussion is incomplete and that the propositions cannot adequately describe all of the
subtieness of technological development as it might actually unfold. Rather, our intent here is
First, we propose the theory begin with the assumption that technology is, in essence, a
body of knowledge. 5 Even though the ultimate goal may be to produce something, the
currency of R&D communities is not so much actual things as it is the ideas, or theories, about
how and why things work the way they do. Therefore, technological development can be
^The idea that technology is essentially knowledge has gained acceptance among scholars in several disciplines.
For example, see Arrow (1962), Layton (1974), and Constant (1980). Constant, in particular, has pioneered the effort
to understand technological communities with his study of the turbojet. The work of Kaiz and Allen (1985) focuses on
technology development as an information processing and problem-solving activity. The intent here is to link the
concepts of information, knowledge, and problem-solving into a unified model of technological development.
31
understood as an intellectual process that evolves over time, whereby new knowledge is
created and applied in order to construct a new product or process. Of course, emerging
technologies are not developed from scratch but are a combination of newly created knowledge
and existing knowledge drawn from other epistemic realms. However, in the early stage of a
technology's emergence, this body of knowledge is incomplete— one simply does not know
everything one needs to know in order to make the technology work. The areas in which
successfully reduced to practice, then new knowledge, in the form of solutions to the
The central actors in this process are the individual researchers who become dedicated to
solving the problems, and it is they who set the process in motion with their efforts to create
and apply knowledge. In the course of their work, researchers perform three basic activities:
(a) they produce information, (b) they transform information into knowledge, or in other
words, they solve problems, and (c) they communicate information and knowledge to each
other. We propose that the rate of progress in a technology's emergence is a function of how
quickly problems are solved, which, in turn, depends on the amount of information produced,
the number of diverse solutions attempted, and the extent to which information and knowledge
researcher, the more likely he is to arrive at a useful solution. Moreover, we anticipate that the
more diversity in the types of solutions attempted, the more likely that critical solutions will be
It is reasonable to assume that researchers are rational, in the economic sense that they are
motivated by self-interest: that is, they are eager to solve problems because there are rewards
32
for those who do. The researcher's objective is to maximize the amount of knowledge he
produces and can lay claim to before other researchers, because these claims have potential
value. The actual value of a claim is contingent upon whether or not all problems necessary for
reaching commercialization are resolved and the length of time taken. Expressed in another
way, for a given piece of knowledge, it is more valuable in use, and the sooner it is used the
more value it will have. A researcher need not produce all of the knowledge required to
Since his choice is guided by the objective to maximize knowledge claims, the researcher is
motivated to participate in those communities, within the realm of his disciplinary expertise, in
which he believes there is a good probability of finding solutions that yield him valuable
claims. If the actual probability is equal to or greater than expected (the task of solving
problems is easier than anticipated), then the researcher will likely remain in the field;
however, if it is lower than expected (solving problems is harder than anticipated), then the
researcher might switch to another community with a higher perceived probability of success.
The switching behavior of a researcher is moderated by the length of time spent contributing to
researcher accumulates over time, the less likely he is to exit a community prior to reaping the
choice, but rather is subject to frequent reconsideration based on new information. The
produced and researchers reevaluate the probabilities of successfully solving the problems they
face. Its attractiveness also may be influenced by changes in other technologies, and indeed a
33
variety of other events, all of which will be reflected in the researcher's decision.
why). Having information does not necessarily imply either.6 The researcher first must make
sense of the information available to him— the cognitive process of transforming information
The processes of information production and transformation are time consuming; therefore,
any individual researcher can only accomplish a certain amount of effort in a given period of
time. The information he produces and the solutions he pursues are an expression of the
researcher's own judgment and creativity, although he may be influenced by people with
whom he collaborates. Unfortunately, not all attempted solutions will work satisfactorily. The
probability that a particular solution will work successfully can only be subjectively
determined a pr/on, and this is likely to change over time as researchers generate new
information (in that it might contribute to the successful solution of a problem) cannot be
determined a priori. For this reason, the value of information is less certain than that of
knowledge.
The amount of information and knowledge available to a researcher depends upon how
^For example, a researcher runs an experiment and obtains negative results. He cannot use the information in the
practical sense that enables him to do or explain something-that is, it is not knowledge. Nevertheless, the
information may be instructive and eventually prove key in creating knowledge. Moreover, even though the
information is not valuable in that it should be published or patented, it does have value in the sense that one who is
much he can produce himself, or receive in the process of communicating with others. The
communication of information and knowledge implies that a researcher can also gather
information and knowledge produced by another researcher, and disseminate to others that
which he produces or learns. For the most part, information is communicated informally by
documented claims, such as with the submission of papers to refereed journals or patent
applications.
organizational boundaries between researchers and their (and their organization's) economic
interests. Organizational boundaries are important for two reasons: first they give rise to
information asymmetries among researchers because they impede the flow of information and
increase the cost of information gathering. As a result, organizational boundaries can slow the
available to each researcher. However, organizational boundaries can also enhance knowledge
production to the extent it increases the diversity of problem solutions pursued by the
community as a whole, since researchers in different organizations are likely to have different
information sets.
To illustrate this point, assume the extreme conditions. In the first case, as the research
community grows, all researchers are employed by the same organization. Thus, all
researchers are exposed to the same information set, but the diversity of solutions performed is
limited by the researchers' mutual influence. In the second case, as the community grows,
each researcher is employed in a separate organization. Thus, each researcher is working fi"om
a different, limited set of information and performing independent solutions, such that the
diversity of attempted solutions is maximized. Simply stated, in one situation each researcher
35
in the community has a wealth of information but a limited number of approaches in solving
the problems being confronted; in the other situation, although the community can generate the
same amount of information, the amount available to any particular researcher is small and the
researchers who act as technological gatekeepers— that is, researchers who tend to communicate
with others in different organizations (Allen, 1979). Given the economically rational behavior
quid pro quo (von Hippel, 1988). Furthermore, information is more likely to be the trading
object of grapevines than is knowledge, because the value of information is indeterminate and
because knowledge (which does have potential value) needs to be formally documented when
The propositions presented in this discussion, taken together, suggest that R&D
communities are freely functioning, loosely connected coalitions of researchers, who in their
search for promising opponunities to expand on the frontier of knowledge, become dedicated
to developing a common set of ideas. It is a self-organizing process, which grows out of the
efforts of researchers who endeavor to solve interesting problems as they lock themselves into
a struggle with nature. As researchers probe the frontier, sometimes they may find it stubborn
and unyielding, and in other instances they may uncover an area which gives way to rich
The speed at which a promising technology develops depends on how rapidly a community
forms and grows, and the structural and behavioral characteristics the community adopts as it
matures. Some characteristics promote cooperation between researchers, while others foster
36
competition. In combination, both elements are important contributors to the swift conduct of
the problem-solving process: the competition that breeds with the rapid growth and diffusion
of the community, escalating the amount of information produced and increasing the diversity
of solutions pursued; and the cooperation which flows from the grapevine, enabling
researchers to exchange notes on the latest experiments and to pursue solutions without
Conclusion
The objective of this research is to draw attention to the notion that R&D communities exist
and that they may play an instrumental role in the emergence of new technologies. Although
the importance of academic communities in the operation of science is generally accepted, very
little is understood about how communities may contribute to technological progress. This
continues to be the case, despite the fact that it is fairly common to hear people talk of entities,
such as the "software development community," which are so intimately involved with
communities function? Can their internal operation tell us anything about the rate at which
empirically assessed the growth and diffusion of the community using the literature published
by neural networks researchers as a source of information. The data indicate that the neural
network community emerged with great speed, indicative of a bandwagon-effect, but only after
experiencing many years of wavering enthusiasm and despair. The community survived
intense controversy and lack of financial support, to persevere on the dedication of a small
number of researchers who eventually succeeded in advancing their ideas to a state whereby
37
they began to attract the commitment of others. As it grew rapidly, the community also became
widely dispersed across hundreds of organizations in the public and private sector, and was
held together by a rather elaborate grapevine that preserved the flow of information between
laboratories.
It is unwise to read too much into the present finding without the careful examination of
other R&D communities. However, the emergence of neural networks is suggestive of a self-
organizing process that arises from the grassroots. It is hard to believe, given the evidence,
that the direction of research can be centrally planned. The emergence of a new technology
appears very much to be the result of independent forces, which become unified by a common
commitment to a set of ideas and to solving the problems necessary to make these ideas work.
It also suggests that technological progress might not be accelerated simply by brute force
and organization, Arrow (1974), suggests that "...organizations are a means of achieving the
benefits of collective action in situations in which the price system fails." Following in a
similar vein, it may very well be that R&D communities are a means of achieving the benefits
of collective action in situations in which the single organization fails. Given this perspective,
An organization, such as the firm, can be quite effective in developing a new technology
single-handedly, especially if the choice among problem solutions is narrow. The organization
facilitates the coordination of researchers' activities, reinforces cooperation, and ensures that
the production of information and knowledge is done without duplication and is readily
38
available to all researchers. However, in situations where the choice among problem solutions
competitive pressures to speed the problem solving process. Although communication is more
difficult than within a single organization, it can be improved by the grapevine and the
In the course of this analysis, we have attempted to expose the salient characteristics of
communities, all of which are quite familiar phenomena to the researchers who participate in
them. Indeed, we have drawn largely on the observations and vocabulary of the researchers
themselves in our descriptions of the internal functioning of the community. Many neural
furthermore, they have an intuitive understanding— indeed, only as they could have— of the
controversies, grapevines, and momentum are not rarefied abstractions but well-known
concepts which are interwoven into the fabric of laboratory life. It is difficult to foresee any
comprehensive understanding of emerging technologies that could ignore the role R&D
communities might play. The time is ripe for funher research.
39
References
Allen, Thomas J. (1977) Managing the Flow of Technology. Cambridge, MA: MIT Press.
Allman, William F. (1987) "Designing Computers That Think the Way We Do," Technology Review,
May/June: 59-65.
Constant, Edward W. (1980) The Origins of the Turbojet. Baltimore: Johns Hopkins University Press.
DARPA Neural Networks Study. (1988) Fairfax, VA: AFCEA International Press, 1988.
Hecht-Nielsen, Robert. (1988) "Neurocompuling: Picking the Human Brain," /£££ 5pec<rum 25, 3: 36-41.
Hopfield, John J. (1988) "Artificial Neural Networks," IEEE Circuits and Devices Magazine, Sept.: 3-10.
Johnson, George. (1986) Machinery of the Mind: Inside the New Science of Artificial Intelligence.
Redmond, Wash.: Microsoft Press.
Katz, Ralph and Thomas J. Allen. (1985) "Organizational Issues in the Introduction of New Technologies," in
P.R. Kleindorfer, ed.. The Management of Productivity and Technology in Manufacturing. New York:
Plenum Press.
Kinoshita, June and Nicholas G. Palevsky. (1987) "Computing with Neural Networks," High Technology,
May: 24-31.
Kohonen, Teuvo. (1988) "An Introduction to Neural Computing," Neural Networks 1: 3-16.
Layton, Edward T. (1974) "Technology as Knowledge," Technology and Culture 15: 31-41.
Malina, Debbie. (1988) "Will the Real Intelligent Machine Please Stand Up? Mass High Tech 6, 15: 1, 33.
National Science Foundation. (1988) International Science and Technology Data Update: 1988. (NSF 89- 307)
Washington, DC: Government Printing Office.
Tank, David W. and John J. Hopfield. (1987) "Collective Computation in Neuronlike Circuits," Scientific
Am^ncan, December: 104-114.
von Hippel, Eric A. (1987) "Cooperation Between Rivals: Informal Know-how Trading," Research Policy 16:
291-302.
Widrow, Bernard. (No Date). Letter to member of the International Neural Networks Society.
Ziman, John. (1984) An Introduction to Science Studies. Cambridge: Cambridge University Press.
40
Appendix A
Once upon a time two daugther sciences were bom to the new science of cybernetics. One sister was
natural, with features inherited from the study of the brain, from the way nature does things. The other was
artificial, related from the beginning to the use of computers. Each of the sister sciences tried to build models
of intelligence, but from very difficult materials. The natural sister built hermodels (called neural networks)
out of mathematically purified neurones. The artificial sister built her models out of computer programs.
In their first bloom of youth the two were equally successful and equally pursued by suitors from other
knowledge. They got on very well together. Their relationship changed in the early sixties when a
fields of
new monarch appeared, one with the largest coffers ever seen in the kingdom of the sciences: Lord DARPA, the
Defense Department's Advanced Research Projects Agency. The artificial sister grew jealous and was determined
to keep for herself the access to Lord DARPA's research funds. The natural sister would have to be slain.
The bloody work was attempted by two staunch followers of the artificial sister, Marvin Minsky and
Seymour Papert, cast in the role of the huntsmen sent to slay Snow White and bring back her heart as proof of
the deed. Their weapon was not the dagger but the mightier pen, from which came a book— Perceptrons-
purporting to prove that neural nets could never fill their promise of building models of mind: only computer
programs could do this. Victory seemed assured for the artificial sister. And indeed for the next decade all the
rewards of the kingdom came to progeny, of which the family of expert systems did the best in fame and
fortune.
But Snow White was not dead. What Minsky and Papert had shown the world as proof was not the heart of
the princess; it was the heart of a pig. To be more literal: their book was read as proving that the neural net
approach to building models of the mind was dead. But a closer look reveals that they demonstrated something
less than this. The book did indeed point out very serious limitations of a certain class of nets (nowadays know
as one-layer perceptrons) but was misleading in its suggestion that this class of nets was at the heart of
connectionism.
Connectionist writings present the story as having a happy ending. The natural sister was quietly nurtured
in the laboratories of a few ardent researchers who kept the faith, even when the world at large let itself be
convinced that the enterprise was futile.
From, Seymour Papert, "One AI or Many?" Dcedalus, Volume 117, Number 1, Winter 1988 (p. 4-5)
5960 U I 7
1
Date Due
iNOV 1 1 1^^
Mi ^,
W" 8 1999
OCT" 1 3 ZOO
NOV 1 1 zoqa
liillllilllill
3 TDflO D0S70M3M fl